Inspiration

Misinformation spreads fastest in private group chats. Well-meaning relatives forward "urgent" health tips, microchip-vaccine claims, and fake news screenshots because they want to protect the people they love. Checking each message takes time and skill, and it's even harder when the hoax isn't in English. We wanted to make fact-checking as easy as forwarding.

What it does

FactSnap checks a pasted message, a news link, or a screenshot in seconds. It breaks the content into specific claims, finds evidence from trusted fact-checkers, and gives each claim a verdict (True, False, Misleading, or Unverified) with real citations. It also flags manipulation tactics like clickbait and pressure to forward, and shows an overall credibility score only after every claim has been judged. Then it writes a short, polite reply in the same language as the original message, ready to paste back into the group. Any language in, the same language out.

How we built it

FactSnap is a Next.js 15 app on Vercel. DeepSeek V4.1 Flash, served on Fireworks AI, reads text and images natively, so no separate OCR step is needed. It extracts the claims, and Exa searches the web for each one, prioritizing outlets like Reuters, AFP, Snopes and local fact-checkers. A separate judging step evaluates each claim using only the sources Exa actually retrieved, which keeps verdicts grounded and stops the model from inventing citations. Every step streams to the UI live as NDJSON, so users can see how the conclusion was reached.

Challenges we ran into

  • Keeping verdicts grounded: the judge only sees retrieved sources and must say "Unverified" when the evidence is missing.
  • Multilingual hoaxes: detecting the input language and making sure the search, the verdicts and the reply all stay in that language.
  • Streaming UX: partial results looked like final verdicts at first, so we added a neutral "Checking…" state until every claim is judged.
  • Staying fast and free on serverless limits.

Accomplishments that we're proud of

  • A full check runs in about 5–15 seconds, with sources a user can click and verify.
  • It works on screenshots in languages like Indonesian, which is how hoaxes actually spread.
  • The copy-paste reply turns a fact-check into something people will actually send.

What we learned

Grounding matters more than model size. Restricting the judge to retrieved evidence made the results far more trustworthy. We also learned that the hardest part of fighting misinformation is the social side, so the friendly reply matters as much as the verdict.

What's next for FactSnap

  • A WhatsApp and Telegram bot, so people can forward a message and get a check back
  • A browser extension
  • Caching checks for viral hoaxes we've already seen
  • Support for audio and video forwards

Built With

  • deepseek
  • next.js
Share this project:

Updates

Submission history